cat RAPID omics : a web server for large-scale prediction of protein–RNA interactions

1Increasing evidence indicates that ribonucleoprotein interactions are fundamental for cellular regulation. Moreover, several studies highlighted the involvement of RNA molecules in the onset and progression of human diseases including neurological disorders. To our knowledge, there are two sequence-based methods for prediction of protein–RNA interactions: catRAPID and RPISeq. The catRAPID algorithm exploits predictions of secondary structure, hydrogen bonding and van der Waals’ contributions to estimate the binding propensity of protein and RNA molecules. RPISeq is based on support vector machine (SVM) and random forest (RF) models predicting protein–RNA interactions from primary structure alone. Both methods show remarkable performances, but catRAPID discriminates positive and negative cases with higher accuracy and has been tested on long non-coding RNAs.

2Here we introduce catRAPID omics to perform high-throughput predictions of protein–RNA interactions using the information on protein and RNA domains involved in macromolecular recognition.

3The catRAPID omics server provides two main services to explore the interaction potential of (i) a protein of interest with respect to a target transcriptome or (ii) a given RNA with respect to the nucleic acid binding proteome. Several options are available to refine the type of analysis in eight model organisms or custom libraries (see online documentation): In the case of a protein query, catRAPID omics takes as input the protein sequence (FASTA format): full-length or, alternatively, nucleic acid binding regions. For a transcript query (FASTA format), the server uses the full-length sequence if below 1200 nt, or, alternatively, uses fragments with predicted stable secondary structure. Full-length proteins and nucleic acid binding regions can be searched. The server automatically detects disordered proteins lacking canonical RNA binding domains. Indeed, it has been observed that disordered regions are enriched in RNA binding proteins.As RNA motifs are important for protein recognition, a search for these elements is carried out. The motifs were taken from RNA-Binding Protein DataBase (RBPDB), SpliceAid-F and a recent motif compendium. Using the interaction propensities distribution, catRAPID omics predicts the RNA binding ability of the input protein (86% accuracy) and ranks RNA interactions (downloadable by the user).

4In the case of a protein query, catRAPID omics takes as input the protein sequence (FASTA format): full-length or, alternatively, nucleic acid binding regions.

5For a transcript query (FASTA format), the server uses the full-length sequence if below 1200 nt, or, alternatively, uses fragments with predicted stable secondary structure. Full-length proteins and nucleic acid binding regions can be searched.

6The server automatically detects disordered proteins lacking canonical RNA binding domains. Indeed, it has been observed that disordered regions are enriched in RNA binding proteins.

7As RNA motifs are important for protein recognition, a search for these elements is carried out. The motifs were taken from RNA-Binding Protein DataBase (RBPDB), SpliceAid-F and a recent motif compendium.

8Using the interaction propensities distribution, catRAPID omics predicts the RNA binding ability of the input protein (86% accuracy) and ranks RNA interactions (downloadable by the user).

9In the output page (Fig. 1A), we report all the variables used to estimate protein–RNA associations: interaction propensity, discriminative power, interaction strength and presence of protein RNA binding domains as well as RNA motifs. A ‘star rating system’ ranks the binding propensities (http://service.tartaglialab.com/static_files/shared/faqs.html). As for the reference sets, ENSEMBL (version 68) is used for retrieval and classification of coding and non-coding RNAs, whereas protein sequences are gathered from the UniProtKB database (release 2012_11). Finally, catRAPID omics uses hmmscan, a Hidden Markov Model-based algorithm from the HMMER3 package, to identify known PfamA domains and recognize protein regions involved in binding nucleic acid molecules. Algorithm hit significance is determined according to the PfamA ‘gathering thresholds’.

10The catRAPID algorithm has been previously validated on a number of protein–RNA associations. To evaluate large-scale performances of catRAPID omics, we used data from recent large-scale experiments. To compare predicted and experimental interactions, we used Fisher’s exact test. As shown in Figure 1B, performances on the human splicing factor serine/arginine-rich splicing factor 1 (SRSF1) and murine nucleic acid binding protein Histone-lysine N-methyltransferase EZH2 (EZH2) are highly significant (P-values: 0.01 and 0.01, respectively). Good performances are found for low-throughput experiments on murine non-coding X inactive specific transcript (Xist) repeat A region (RepA) and yeast small nuclear RNA U1 (P-values: 0.03 and 0.015) (Fig. 1B). To illustrate the ability of catRAPID omics to predict interactions with nucleic acid binding domains (Fig. 1C), we used murine FUS and rat TAR DNA-binding protein 43 (TDP43) (P-values: 3e-05 and 0.002) as well as human Crumbs homolog 3 (CRB3) 3′-untranslated region and yeast small nuclear U2 (P-values: 0.001 and 2e-0.6). To evaluate catRAPID’s performances on high-throughput data, we collected positive interactions (TDP43: 568, FUS: 99, SRSF1: 358, EZH2: 1141) as well as negative controls (same numbers as positives and generated in four random extractions). Comparing the interaction scores of positives and negatives, we found enrichment (calculated as discriminative power) in 72% (TDP43), 88% (FUS), 74% (SRSF1) and 56% (EZH2) of cases. On the same datasets, SVM RPIseq showed enrichment in 58% (TDP43; RF has enrichment in 53%), 83% (FUS; RF has enrichment in 68%), 47% (SRSF1; RF has enrichment in 59%) and 41% (EZH2; RF has enrichment in 48%) of cases.

11Despite recent technical developments, detection of protein–RNA associations remains a challenging task. For this reason, we developed an algorithm that can be used to complement experimental efforts. The catRAPID omics server offers unique features such as organism-specific proteomic and transcriptomic libraries, possibility to generate custom datasets, analysis of long sequences and calculation of interaction specificities. Moreover, we implemented an algorithm for the detection of RNA motifs as well as protein RNA binding domains, which will help to retrieve recognition motifs embedded in sequences. Our server enables fast calculations of ribonucleoprotein associations and predicts RNA binding activity of proteins with high accuracy, thus resulting in a powerful tool for designing new experiments.